Leveraging contextual events on structure-aware next activity prediction
Alessandro Mele, Claudia Diamantini, Domenico Potena
cs.LG · cs.AI
Quick verdict
Proposes four graph encodings of contextual events for next activity prediction; careful 5-log study, but gains are modest and lack significance tests.
Summary
The paper addresses next activity prediction in predictive process monitoring, asking whether explicitly modeling contextual information — events from concurrently running process instances — improves GNN-based prediction. Building on Instance Graphs, the authors derive prefix-Instance Graphs and propose four encoding strategies: E1/E2 attach contextual event nodes directly to the prefix graph (differing in attachment point), while E3/E4 keep a separate, connected context graph, with E4 including the complete contextual prefix-IGs. Node features combine one-hot activity encoding with three temporal features; a GraphSAGE-style spatial GNN with readout, MLP, and softmax performs graph classification. On five public event logs (BPI12WC, BPI20P, BPI20R, Helpdesk, Receipt), with chronological 67/33 splits, Optuna tuning, and five seeded runs, E4 performs best overall but only marginally above E3, while E2 is clearly worst. Against graph-based baselines — the authors' prior IG approach [2], its inter-case feature extension [4], and PROPHET [13] — E4 improves accuracy by average margins of 2.70% and 7.75% respectively, leading all 15 metric-dataset cells in Table 4. Limitations: gains rest on mean/std over five runs without significance testing; context is defined purely by temporal overlap, ignoring resources and causality (acknowledged by the authors); and comparison is restricted to graph methods, omitting sequence models. Code and experiment details are released in a public GitHub repository.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Encoding contextual process instances as a separate connected context graph (E3/E4) yields better next activity prediction than attaching contextual event nodes to the prefix graph (E1/E2).
- Strategy E4, which incorporates complete contextual prefix-Instance Graphs, achieves the best overall performance, though only marginally above E3.
- The selected strategy outperforms graph-based baselines [2], [4], and [13] across five real-world event logs, with average accuracy gains of 2.70% over [4] and 7.75% over [13].
- Incorporating contextual process instances benefits next activity prediction compared with structure-only Instance Graph approaches.
Strengths
Systematic encoding-strategy ablation
Four encoding strategies are compared head-to-head across five datasets with five seeded runs each (Table 3), directly answering RQ1. The finding that separating prefix and context graphs (E3/E4) beats attaching context nodes to the prefix graph (E1/E2) — with E2's tightest coupling worst by an average 4.95% Acc against E1 — is an actionable design insight for graph-based PPM.
Rigorous leakage-prevention protocol
Section 4 details controls rarely stated this explicitly: the process model is mined only from the training set, feature normalization uses training maxima, and contextual events are restricted to the same data split as the prefix, with the induced train/test-boundary sparsity openly discussed rather than hidden.
Honest, granular reporting
Reproducibility provision
Weaknesses
No statistical significance testing
Modest gain over the closest baseline, which is the authors' own prior work
Baseline scope restricted to graph methods
Unreported tuning budget and final configurations
E1-vs-E2 gap left unexplained
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets next activity prediction within predictive process monitoring (PPM) and asks whether explicitly encoding contextual process instances — events from concurrently running cases — improves GNN-based predictors. Central claims: (i) of four proposed encoding strategies, separating the prefix graph from a connected context graph (E3/E4) outperforms attaching contextual nodes to the prefix graph (E1/E2); (ii) E4, which embeds complete contextual prefix-Instance Graphs, is best overall though only marginally above E3; (iii) the selected strategy beats three graph-based competitors [2], [4], [13] on five public logs. RQ1 is answered via Table 3, RQ2 via Table 4.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is contextual next activity prediction?
In predictive process monitoring, next activity prediction forecasts the next activity a running business process case will execute, based on its event history so far. The contextual variant additionally exploits information from other cases running at the same time (e.g., concurrent events or system load), because such conditions can change how a case evolves — for example, a help desk ticket may be delayed when operators are busy with many other tickets.
Which encoding strategy works best in this paper?
Strategy E4 — building a separate connected context graph that contains the complete prefix-Instance Graphs of concurrently running cases — achieved the best overall results across the five event logs, but only marginally better than the simpler E3 (context events only, no full histories). E2, which attaches contextual events to the last node of the prefix graph, performed clearly worst.
Are the code and datasets available for this paper?
Yes. The paper links a public GitHub repository (AlessandroMele/leveraging_contextual_events_nap) with source code and experiment details, and all five event logs — BPI Challenge 2012 (W-complete subprocess), BPI Challenge 2020 (Prepaid Travel Cost and Request for Payment), Helpdesk, and Receipt — are publicly downloadable from the 4TU research data repository.
Does modeling context actually improve next activity prediction?
The paper reports consistent wins over graph-based baselines: average accuracy improvements of 2.70% over aggregated inter-case features (Diamantini et al. 2024) and 7.75% over PROPHET across five logs, with E4 leading every metric-dataset combination in its comparison table. However, results are averages of five runs without statistical significance tests, the gain over the closest baseline is modest, and no sequence models (LSTM/Transformer) were compared.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure content (Figs. 2-5) is garbled in the PDF-to-text extraction (node labels and axis ticks run together, e.g. Fig. 2 'RT E TC AS S'); assessment of figures relies on the surrounding prose, which is complete.
- Table 3 lists E4 Receipt F1-W as '83.6(0.17)' while Table 4 lists '83.66(0.17)' — likely a PDF-extraction artifact of the same value; cited as 83.66 from Table 4.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.